Paper: arXiv 2308.07329

Authors: Avish Buramdoyal, Tim Gebbie

Abstract

Blackjack or “21” is a popular card-based game of chance and skill. The objective of the game is to win by obtaining a hand total higher than the dealer’s without exceeding 21. The ideal blackjack strategy will maximize financial return in the long run while avoiding gambler’s ruin. The stochastic environment and inherent reward structure of blackjack presents an appealing problem to better understand reinforcement learning agents in the presence of environment variations. Here we consider a q-learning solution for optimal play and investigate the rate of learning convergence of the algorithm as a function of deck size. A blackjack simulator allowing for universal blackjack rules is also implemented to demonstrate the extent to which a card counter perfectly using the basic strategy and hi-lo system can bring the house to bankruptcy and how environment variations impact this outcome. The novelty of our work is to place this conceptual understanding of the impact of deck size in the context of learning agent convergence.

Complexity vs Empirical Score

  • Math Complexity: 3.5/10
  • Empirical Rigor: 5.0/10
  • Quadrant: Street Traders — empirical and applied, lower math complexity

Why this score: The paper involves basic game theory and RL (Q-learning) concepts without dense derivations, but implements a simulator and evaluates convergence rates and financial outcomes, making it moderately data and implementation heavy.

Research Flowchart

  flowchart TD
  subgraph R ["Research Goal"]
      G["How does deck size affect Q-learning convergence and card counting effectiveness in Blackjack?"]
  end

  subgraph M ["Methodology"]
      S1["Implement Blackjack Simulator"] --> S2["Develop Q-learning Agent"]
      S2 --> S3["Test varying Deck Sizes"]
      S3 --> S4["Simulate Hi-Lo Card Counter"]
  end

  subgraph D ["Data & Inputs"]
      D1["Deck Sizes: 1 to 8 Decks"]
      D2["Standard Blackjack Rules"]
      D3["Q-learning Parameters"]
  end

  subgraph P ["Computational Process"]
      P1["Train Agent: Update Q-values via rewards"]
      P2["Evaluate Win Rate & Convergence Speed"]
      P3["Simulate Bankroll: Counter vs House Edge"]
  end

  subgraph O ["Key Findings/Outcomes"]
      F1["Q-learning converges faster with fewer decks"]
      F2["Card counting significantly reduces house edge"]
      F3["Deck size impacts optimal strategy effectiveness"]
  end

  G --> S1
  D1 & D2 & D3 --> P1
  P1 --> P2 --> P3
  P3 --> F1 & F2 & F3

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes a Q‑learning based agent for optimal blackjack play and investigates how the rate of learning convergence varies with deck size.

Source sentence
Here we consider a q-learning solution for optimal play and investigate the rate of learning convergence of the algorithm as a function of deck size. The novelty of our work is to place this conceptual understanding of the impact of deck size in the context of learning agent convergence.

Assumes 6

  • Rewards are not discounted (γ = 1) and terminal rewards represent returns. model
    Source sentence
    We note that rewards γ are not discounted, i.e. γ = 1 and so terminal rewards represent returns.
  • The state is defined by the player’s hand total and the dealer’s face‑up card. model
    Source sentence
    The state depends on the player’s hand total and the dealer’s face up card.
  • Episodes begin with a randomly chosen state and action (exploring starts). model
    Source sentence
    Episodes are therefore generated with exploring starts. This means that each episode begins with a randomly chosen state and action and then follows the current policy.
  • A decaying epsilon greedy policy is used to balance exploration and exploitation. model
    Source sentence
    F. Q-Learning agent: Hit and stand only For the base q-learning model, we implement a decaying epsilon q-learning algorithm, using the one step q-learning given by Equation 9 for the blackjack environment assumed by ”blackjack-v0”. To ensure an effective learning process, appropriate values for ϵ. ϵ is a hyperparameter used to control the balance between exploration and exploitation in many reinforcement learning
Show 2 more
  • The dealer may either stand or hit on a soft 17, as a configurable rule. model
    Source sentence
    We allow the dealer the choice to either stand or hit on having a soft 17.
  • The card counter uses the Hi‑Low counting method and follows basic strategy perfectly. model
    Source sentence
    The card counter uses the Hi-Low method and the other players randomly decide on the % of their bankroll to bet for each round.

Data

  • Blackjack simulation data (random agents, card counter, dealer) synthetic public
    Source sentence
    The figure can be recovered using python script random agent.py on GitHub resource [23]. Case 1: 4 Players, 6 Decks, Dealer stands on soft 17 and 10000 simulations Players Win % Draw % Loss % Bankroll Card counter 14.55 % 1.35 % 3.21 % R14, 647, 394 Random agent 1 11.03 % 1.43 % 7.52 % -R6, 835, 650 Random agent 2 11.01 % 1.44 % 7.52 % -R6, 989, 950 Random agent 3 11.02 % 1

Costs

Not applicable The paper reports no trading or execution result.

Finds 5

  • The random agent achieves an average payoff of –396.29. performance
    Source sentence
    The average payoff of -396.29 achieved by the random agent is to be compared to the payoff achieved by the q-learning agent for following q-learning policy.
  • The card counter outperforms random agents across simulation sizes, numbers of players, and deck sizes. performance
    Source sentence
    1) Performance across simulation size: Figure 8 shows a clear outperformance of the card counter to the random agent against the dealer across both variations in the number of players, deck size and whether the dealer hits or stands on a soft 17 hand. 2) Performance across number of players: Figure 9 also indicates the outperformance of the card counter to the random agent across increased number of players and
  • The number of decks in play does not influence the win‑percentage discrepancy between the card counter and random agents. performance
    Source sentence
    The number of decks in play do not seem to influence the discrepancy in win% against the dealer between the card counter and the random agent.
  • Even with an initial bankroll ratio of 1:200 favoring the house, the dealer goes bankrupt when the card counter uses perfect basic strategy and the Hi‑Lo system. performance
    Source sentence
    Even for an initial bankroll ratio of 1:200 in favour of the house, the dealer goes bankrupt for having a card counter perfectly using the basic strategy table and the Hi-Lo system.
  • Learning convergence rate of the Q‑learning algorithm varies as a function of deck size. theoretical
    Source sentence
    Here we consider a q-learning solution for optimal play and investigate the rate of learning convergence of the algorithm as a function of deck size.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 15 related papers). Coverage is partial, so this is neither replication nor refutation.

Code

No repository by the authors found in the paper.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.